1 //===------- llvm/CodeGen/ScheduleDAG.h - Common Base Class------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the ScheduleDAG class, which is used as the common
11 // base class for instruction schedulers.
13 //===----------------------------------------------------------------------===//
15 #ifndef LLVM_CODEGEN_SCHEDULEDAG_H
16 #define LLVM_CODEGEN_SCHEDULEDAG_H
18 #include "llvm/CodeGen/MachineBasicBlock.h"
19 #include "llvm/Target/TargetMachine.h"
20 #include "llvm/ADT/DenseMap.h"
21 #include "llvm/ADT/BitVector.h"
22 #include "llvm/ADT/GraphTraits.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/PointerIntPair.h"
29 class MachineConstantPool;
30 class MachineFunction;
31 class MachineRegisterInfo;
33 class TargetRegisterInfo;
36 class TargetInstrInfo;
37 class TargetInstrDesc;
39 class TargetRegisterClass;
40 template<class Graph> class GraphWriter;
42 /// SDep - Scheduling dependency. This represents one direction of an
43 /// edge in the scheduling DAG.
46 /// Kind - These are the different kinds of scheduling dependencies.
48 Data, ///< Regular data dependence (aka true-dependence).
49 Anti, ///< A register anti-dependedence (aka WAR).
50 Output, ///< A register output-dependence (aka WAW).
51 Order ///< Any other ordering dependency.
55 /// Dep - A pointer to the depending/depended-on SUnit, and an enum
56 /// indicating the kind of the dependency.
57 PointerIntPair<SUnit *, 2, Kind> Dep;
59 /// Contents - A union discriminated by the dependence kind.
61 /// Reg - For Data, Anti, and Output dependencies, the associated
62 /// register. For Data dependencies that don't currently have a register
63 /// assigned, this is set to zero.
66 /// Order - Additional information about Order dependencies.
68 /// isNormalMemory - True if both sides of the dependence
69 /// access memory in non-volatile and fully modeled ways.
70 bool isNormalMemory : 1;
72 /// isMustAlias - True if both sides of the dependence are known to
73 /// access the same memory.
76 /// isArtificial - True if this is an artificial dependency, meaning
77 /// it is not necessary for program correctness, and may be safely
78 /// deleted if necessary.
79 bool isArtificial : 1;
83 /// Latency - The time associated with this edge. Often this is just
84 /// the value of the Latency field of the predecessor, however advanced
85 /// models may provide additional information about specific edges.
89 /// SDep - Construct a null SDep. This is only for use by container
90 /// classes which require default constructors. SUnits may not
91 /// have null SDep edges.
92 SDep() : Dep(0, Data) {}
94 /// SDep - Construct an SDep with the specified values.
95 SDep(SUnit *S, Kind kind, unsigned latency = 1, unsigned Reg = 0,
96 bool isNormalMemory = false, bool isMustAlias = false,
97 bool isArtificial = false)
98 : Dep(S, kind), Contents(), Latency(latency) {
103 "SDep::Anti and SDep::Output must use a non-zero Reg!");
106 assert(!isMustAlias && "isMustAlias only applies with SDep::Order!");
107 assert(!isArtificial && "isArtificial only applies with SDep::Order!");
111 assert(Reg == 0 && "Reg given for non-register dependence!");
112 Contents.Order.isNormalMemory = isNormalMemory;
113 Contents.Order.isMustAlias = isMustAlias;
114 Contents.Order.isArtificial = isArtificial;
119 bool operator==(const SDep &Other) const {
120 if (Dep != Other.Dep || Latency != Other.Latency) return false;
121 switch (Dep.getInt()) {
125 return Contents.Reg == Other.Contents.Reg;
127 return Contents.Order.isNormalMemory ==
128 Other.Contents.Order.isNormalMemory &&
129 Contents.Order.isMustAlias == Other.Contents.Order.isMustAlias &&
130 Contents.Order.isArtificial == Other.Contents.Order.isArtificial;
132 assert(0 && "Invalid dependency kind!");
136 bool operator!=(const SDep &Other) const {
137 return !operator==(Other);
140 /// getLatency - Return the latency value for this edge, which roughly
141 /// means the minimum number of cycles that must elapse between the
142 /// predecessor and the successor, given that they have this edge
144 unsigned getLatency() const {
148 /// setLatency - Set the latency for this edge.
149 void setLatency(unsigned Lat) {
153 //// getSUnit - Return the SUnit to which this edge points.
154 SUnit *getSUnit() const {
155 return Dep.getPointer();
158 //// setSUnit - Assign the SUnit to which this edge points.
159 void setSUnit(SUnit *SU) {
163 /// getKind - Return an enum value representing the kind of the dependence.
164 Kind getKind() const {
168 /// isCtrl - Shorthand for getKind() != SDep::Data.
169 bool isCtrl() const {
170 return getKind() != Data;
173 /// isNormalMemory - Test if this is an Order dependence between two
174 /// memory accesses where both sides of the dependence access memory
175 /// in non-volatile and fully modeled ways.
176 bool isNormalMemory() const {
177 return getKind() == Order && Contents.Order.isNormalMemory;
180 /// isMustAlias - Test if this is an Order dependence that is marked
181 /// as "must alias", meaning that the SUnits at either end of the edge
182 /// have a memory dependence on a known memory location.
183 bool isMustAlias() const {
184 return getKind() == Order && Contents.Order.isMustAlias;
187 /// isArtificial - Test if this is an Order dependence that is marked
188 /// as "artificial", meaning it isn't necessary for correctness.
189 bool isArtificial() const {
190 return getKind() == Order && Contents.Order.isArtificial;
193 /// isAssignedRegDep - Test if this is a Data dependence that is
194 /// associated with a register.
195 bool isAssignedRegDep() const {
196 return getKind() == Data && Contents.Reg != 0;
199 /// getReg - Return the register associated with this edge. This is
200 /// only valid on Data, Anti, and Output edges. On Data edges, this
201 /// value may be zero, meaning there is no associated register.
202 unsigned getReg() const {
203 assert((getKind() == Data || getKind() == Anti || getKind() == Output) &&
204 "getReg called on non-register dependence edge!");
208 /// setReg - Assign the associated register for this edge. This is
209 /// only valid on Data, Anti, and Output edges. On Anti and Output
210 /// edges, this value must not be zero. On Data edges, the value may
211 /// be zero, which would mean that no specific register is associated
213 void setReg(unsigned Reg) {
214 assert((getKind() == Data || getKind() == Anti || getKind() == Output) &&
215 "setReg called on non-register dependence edge!");
216 assert((getKind() != Anti || Reg != 0) &&
217 "SDep::Anti edge cannot use the zero register!");
218 assert((getKind() != Output || Reg != 0) &&
219 "SDep::Output edge cannot use the zero register!");
224 /// SUnit - Scheduling unit. This is a node in the scheduling DAG.
227 SDNode *Node; // Representative node.
228 MachineInstr *Instr; // Alternatively, a MachineInstr.
230 SUnit *OrigNode; // If not this, the node from which
231 // this node was cloned.
233 // Preds/Succs - The SUnits before/after us in the graph. The boolean value
234 // is true if the edge is a token chain edge, false if it is a value edge.
235 SmallVector<SDep, 4> Preds; // All sunit predecessors.
236 SmallVector<SDep, 4> Succs; // All sunit successors.
238 typedef SmallVector<SDep, 4>::iterator pred_iterator;
239 typedef SmallVector<SDep, 4>::iterator succ_iterator;
240 typedef SmallVector<SDep, 4>::const_iterator const_pred_iterator;
241 typedef SmallVector<SDep, 4>::const_iterator const_succ_iterator;
243 unsigned NodeNum; // Entry # of node in the node vector.
244 unsigned NodeQueueId; // Queue id of node.
245 unsigned short Latency; // Node latency.
246 unsigned NumPreds; // # of SDep::Data preds.
247 unsigned NumSuccs; // # of SDep::Data sucss.
248 unsigned NumPredsLeft; // # of preds not scheduled.
249 unsigned NumSuccsLeft; // # of succs not scheduled.
250 bool isTwoAddress : 1; // Is a two-address instruction.
251 bool isCommutable : 1; // Is a commutable instruction.
252 bool hasPhysRegDefs : 1; // Has physreg defs that are being used.
253 bool hasPhysRegClobbers : 1; // Has any physreg defs, used or not.
254 bool isPending : 1; // True once pending.
255 bool isAvailable : 1; // True once available.
256 bool isScheduled : 1; // True once scheduled.
257 bool isScheduleHigh : 1; // True if preferable to schedule high.
258 bool isCloned : 1; // True if this node has been cloned.
259 Sched::Preference SchedulingPref; // Scheduling preference.
261 SmallVector<MachineInstr*, 4> DbgInstrList; // dbg_values referencing this.
263 bool isDepthCurrent : 1; // True if Depth is current.
264 bool isHeightCurrent : 1; // True if Height is current.
265 unsigned Depth; // Node depth.
266 unsigned Height; // Node height.
268 const TargetRegisterClass *CopyDstRC; // Is a special copy node if not null.
269 const TargetRegisterClass *CopySrcRC;
271 /// SUnit - Construct an SUnit for pre-regalloc scheduling to represent
272 /// an SDNode and any nodes flagged to it.
273 SUnit(SDNode *node, unsigned nodenum)
274 : Node(node), Instr(0), OrigNode(0), NodeNum(nodenum),
275 NodeQueueId(0), Latency(0), NumPreds(0), NumSuccs(0), NumPredsLeft(0),
276 NumSuccsLeft(0), isTwoAddress(false), isCommutable(false),
277 hasPhysRegDefs(false), hasPhysRegClobbers(false),
278 isPending(false), isAvailable(false), isScheduled(false),
279 isScheduleHigh(false), isCloned(false),
280 SchedulingPref(Sched::None),
281 isDepthCurrent(false), isHeightCurrent(false), Depth(0), Height(0),
282 CopyDstRC(NULL), CopySrcRC(NULL) {}
284 /// SUnit - Construct an SUnit for post-regalloc scheduling to represent
286 SUnit(MachineInstr *instr, unsigned nodenum)
287 : Node(0), Instr(instr), OrigNode(0), NodeNum(nodenum),
288 NodeQueueId(0), Latency(0), NumPreds(0), NumSuccs(0), NumPredsLeft(0),
289 NumSuccsLeft(0), isTwoAddress(false), isCommutable(false),
290 hasPhysRegDefs(false), hasPhysRegClobbers(false),
291 isPending(false), isAvailable(false), isScheduled(false),
292 isScheduleHigh(false), isCloned(false),
293 SchedulingPref(Sched::None),
294 isDepthCurrent(false), isHeightCurrent(false), Depth(0), Height(0),
295 CopyDstRC(NULL), CopySrcRC(NULL) {}
297 /// SUnit - Construct a placeholder SUnit.
299 : Node(0), Instr(0), OrigNode(0), NodeNum(~0u),
300 NodeQueueId(0), Latency(0), NumPreds(0), NumSuccs(0), NumPredsLeft(0),
301 NumSuccsLeft(0), isTwoAddress(false), isCommutable(false),
302 hasPhysRegDefs(false), hasPhysRegClobbers(false),
303 isPending(false), isAvailable(false), isScheduled(false),
304 isScheduleHigh(false), isCloned(false),
305 SchedulingPref(Sched::None),
306 isDepthCurrent(false), isHeightCurrent(false), Depth(0), Height(0),
307 CopyDstRC(NULL), CopySrcRC(NULL) {}
309 /// setNode - Assign the representative SDNode for this SUnit.
310 /// This may be used during pre-regalloc scheduling.
311 void setNode(SDNode *N) {
312 assert(!Instr && "Setting SDNode of SUnit with MachineInstr!");
316 /// getNode - Return the representative SDNode for this SUnit.
317 /// This may be used during pre-regalloc scheduling.
318 SDNode *getNode() const {
319 assert(!Instr && "Reading SDNode of SUnit with MachineInstr!");
323 /// setInstr - Assign the instruction for the SUnit.
324 /// This may be used during post-regalloc scheduling.
325 void setInstr(MachineInstr *MI) {
326 assert(!Node && "Setting MachineInstr of SUnit with SDNode!");
330 /// getInstr - Return the representative MachineInstr for this SUnit.
331 /// This may be used during post-regalloc scheduling.
332 MachineInstr *getInstr() const {
333 assert(!Node && "Reading MachineInstr of SUnit with SDNode!");
337 /// addPred - This adds the specified edge as a pred of the current node if
338 /// not already. It also adds the current node as a successor of the
340 void addPred(const SDep &D);
342 /// removePred - This removes the specified edge as a pred of the current
343 /// node if it exists. It also removes the current node as a successor of
344 /// the specified node.
345 void removePred(const SDep &D);
347 /// getDepth - Return the depth of this node, which is the length of the
348 /// maximum path up to any node with has no predecessors.
349 unsigned getDepth() const {
351 const_cast<SUnit *>(this)->ComputeDepth();
355 /// getHeight - Return the height of this node, which is the length of the
356 /// maximum path down to any node with has no successors.
357 unsigned getHeight() const {
358 if (!isHeightCurrent)
359 const_cast<SUnit *>(this)->ComputeHeight();
363 /// setDepthToAtLeast - If NewDepth is greater than this node's
364 /// depth value, set it to be the new depth value. This also
365 /// recursively marks successor nodes dirty.
366 void setDepthToAtLeast(unsigned NewDepth);
368 /// setDepthToAtLeast - If NewDepth is greater than this node's
369 /// depth value, set it to be the new height value. This also
370 /// recursively marks predecessor nodes dirty.
371 void setHeightToAtLeast(unsigned NewHeight);
373 /// setDepthDirty - Set a flag in this node to indicate that its
374 /// stored Depth value will require recomputation the next time
375 /// getDepth() is called.
376 void setDepthDirty();
378 /// setHeightDirty - Set a flag in this node to indicate that its
379 /// stored Height value will require recomputation the next time
380 /// getHeight() is called.
381 void setHeightDirty();
383 /// isPred - Test if node N is a predecessor of this node.
384 bool isPred(SUnit *N) {
385 for (unsigned i = 0, e = (unsigned)Preds.size(); i != e; ++i)
386 if (Preds[i].getSUnit() == N)
391 /// isSucc - Test if node N is a successor of this node.
392 bool isSucc(SUnit *N) {
393 for (unsigned i = 0, e = (unsigned)Succs.size(); i != e; ++i)
394 if (Succs[i].getSUnit() == N)
399 void dump(const ScheduleDAG *G) const;
400 void dumpAll(const ScheduleDAG *G) const;
401 void print(raw_ostream &O, const ScheduleDAG *G) const;
405 void ComputeHeight();
408 //===--------------------------------------------------------------------===//
409 /// SchedulingPriorityQueue - This interface is used to plug different
410 /// priorities computation algorithms into the list scheduler. It implements
411 /// the interface of a standard priority queue, where nodes are inserted in
412 /// arbitrary order and returned in priority order. The computation of the
413 /// priority and the representation of the queue are totally up to the
414 /// implementation to decide.
416 class SchedulingPriorityQueue {
419 SchedulingPriorityQueue() : CurCycle(0) {}
420 virtual ~SchedulingPriorityQueue() {}
422 virtual void initNodes(std::vector<SUnit> &SUnits) = 0;
423 virtual void addNode(const SUnit *SU) = 0;
424 virtual void updateNode(const SUnit *SU) = 0;
425 virtual void releaseState() = 0;
427 virtual bool empty() const = 0;
428 virtual void push(SUnit *U) = 0;
430 void push_all(const std::vector<SUnit *> &Nodes) {
431 for (std::vector<SUnit *>::const_iterator I = Nodes.begin(),
432 E = Nodes.end(); I != E; ++I)
436 virtual SUnit *pop() = 0;
438 virtual void remove(SUnit *SU) = 0;
440 /// ScheduledNode - As each node is scheduled, this method is invoked. This
441 /// allows the priority function to adjust the priority of related
442 /// unscheduled nodes, for example.
444 virtual void ScheduledNode(SUnit *) {}
446 virtual void UnscheduledNode(SUnit *) {}
448 void setCurCycle(unsigned Cycle) {
452 unsigned getCurCycle() const {
459 MachineBasicBlock *BB; // The block in which to insert instructions
460 MachineBasicBlock::iterator InsertPos;// The position to insert instructions
461 const TargetMachine &TM; // Target processor
462 const TargetInstrInfo *TII; // Target instruction information
463 const TargetRegisterInfo *TRI; // Target processor register info
464 MachineFunction &MF; // Machine function
465 MachineRegisterInfo &MRI; // Virtual/real register map
466 std::vector<SUnit*> Sequence; // The schedule. Null SUnit*'s
467 // represent noop instructions.
468 std::vector<SUnit> SUnits; // The scheduling units.
469 SUnit EntrySU; // Special node for the region entry.
470 SUnit ExitSU; // Special node for the region exit.
472 explicit ScheduleDAG(MachineFunction &mf);
474 virtual ~ScheduleDAG();
476 /// viewGraph - Pop up a GraphViz/gv window with the ScheduleDAG rendered
481 /// EmitSchedule - Insert MachineInstrs into the MachineBasicBlock
482 /// according to the order specified in Sequence.
484 virtual MachineBasicBlock *EmitSchedule() = 0;
486 void dumpSchedule() const;
488 virtual void dumpNode(const SUnit *SU) const = 0;
490 /// getGraphNodeLabel - Return a label for an SUnit node in a visualization
491 /// of the ScheduleDAG.
492 virtual std::string getGraphNodeLabel(const SUnit *SU) const = 0;
494 /// addCustomGraphFeatures - Add custom features for a visualization of
496 virtual void addCustomGraphFeatures(GraphWriter<ScheduleDAG*> &) const {}
499 /// VerifySchedule - Verify that all SUnits were scheduled and that
500 /// their state is consistent.
501 void VerifySchedule(bool isBottomUp);
505 /// Run - perform scheduling.
507 void Run(MachineBasicBlock *bb, MachineBasicBlock::iterator insertPos);
509 /// BuildSchedGraph - Build SUnits and set up their Preds and Succs
510 /// to form the scheduling dependency graph.
512 virtual void BuildSchedGraph(AliasAnalysis *AA) = 0;
514 /// ComputeLatency - Compute node latency.
516 virtual void ComputeLatency(SUnit *SU) = 0;
518 /// ComputeOperandLatency - Override dependence edge latency using
519 /// operand use/def information
521 virtual void ComputeOperandLatency(SUnit *, SUnit *,
524 /// Schedule - Order nodes according to selected style, filling
525 /// in the Sequence member.
527 virtual void Schedule() = 0;
529 /// ForceUnitLatencies - Return true if all scheduling edges should be given
530 /// a latency value of one. The default is to return false; schedulers may
531 /// override this as needed.
532 virtual bool ForceUnitLatencies() const { return false; }
534 /// EmitNoop - Emit a noop instruction.
538 void EmitPhysRegCopy(SUnit *SU, DenseMap<SUnit*, unsigned> &VRBaseMap);
541 class SUnitIterator : public std::iterator<std::forward_iterator_tag,
546 SUnitIterator(SUnit *N, unsigned Op) : Node(N), Operand(Op) {}
548 bool operator==(const SUnitIterator& x) const {
549 return Operand == x.Operand;
551 bool operator!=(const SUnitIterator& x) const { return !operator==(x); }
553 const SUnitIterator &operator=(const SUnitIterator &I) {
554 assert(I.Node==Node && "Cannot assign iterators to two different nodes!");
559 pointer operator*() const {
560 return Node->Preds[Operand].getSUnit();
562 pointer operator->() const { return operator*(); }
564 SUnitIterator& operator++() { // Preincrement
568 SUnitIterator operator++(int) { // Postincrement
569 SUnitIterator tmp = *this; ++*this; return tmp;
572 static SUnitIterator begin(SUnit *N) { return SUnitIterator(N, 0); }
573 static SUnitIterator end (SUnit *N) {
574 return SUnitIterator(N, (unsigned)N->Preds.size());
577 unsigned getOperand() const { return Operand; }
578 const SUnit *getNode() const { return Node; }
579 /// isCtrlDep - Test if this is not an SDep::Data dependence.
580 bool isCtrlDep() const {
581 return getSDep().isCtrl();
583 bool isArtificialDep() const {
584 return getSDep().isArtificial();
586 const SDep &getSDep() const {
587 return Node->Preds[Operand];
591 template <> struct GraphTraits<SUnit*> {
592 typedef SUnit NodeType;
593 typedef SUnitIterator ChildIteratorType;
594 static inline NodeType *getEntryNode(SUnit *N) { return N; }
595 static inline ChildIteratorType child_begin(NodeType *N) {
596 return SUnitIterator::begin(N);
598 static inline ChildIteratorType child_end(NodeType *N) {
599 return SUnitIterator::end(N);
603 template <> struct GraphTraits<ScheduleDAG*> : public GraphTraits<SUnit*> {
604 typedef std::vector<SUnit>::iterator nodes_iterator;
605 static nodes_iterator nodes_begin(ScheduleDAG *G) {
606 return G->SUnits.begin();
608 static nodes_iterator nodes_end(ScheduleDAG *G) {
609 return G->SUnits.end();
613 /// ScheduleDAGTopologicalSort is a class that computes a topological
614 /// ordering for SUnits and provides methods for dynamically updating
615 /// the ordering as new edges are added.
617 /// This allows a very fast implementation of IsReachable, for example.
619 class ScheduleDAGTopologicalSort {
620 /// SUnits - A reference to the ScheduleDAG's SUnits.
621 std::vector<SUnit> &SUnits;
623 /// Index2Node - Maps topological index to the node number.
624 std::vector<int> Index2Node;
625 /// Node2Index - Maps the node number to its topological index.
626 std::vector<int> Node2Index;
627 /// Visited - a set of nodes visited during a DFS traversal.
630 /// DFS - make a DFS traversal and mark all nodes affected by the
631 /// edge insertion. These nodes will later get new topological indexes
632 /// by means of the Shift method.
633 void DFS(const SUnit *SU, int UpperBound, bool& HasLoop);
635 /// Shift - reassign topological indexes for the nodes in the DAG
636 /// to preserve the topological ordering.
637 void Shift(BitVector& Visited, int LowerBound, int UpperBound);
639 /// Allocate - assign the topological index to the node n.
640 void Allocate(int n, int index);
643 explicit ScheduleDAGTopologicalSort(std::vector<SUnit> &SUnits);
645 /// InitDAGTopologicalSorting - create the initial topological
646 /// ordering from the DAG to be scheduled.
647 void InitDAGTopologicalSorting();
649 /// IsReachable - Checks if SU is reachable from TargetSU.
650 bool IsReachable(const SUnit *SU, const SUnit *TargetSU);
652 /// WillCreateCycle - Returns true if adding an edge from SU to TargetSU
653 /// will create a cycle.
654 bool WillCreateCycle(SUnit *SU, SUnit *TargetSU);
656 /// AddPred - Updates the topological ordering to accomodate an edge
657 /// to be added from SUnit X to SUnit Y.
658 void AddPred(SUnit *Y, SUnit *X);
660 /// RemovePred - Updates the topological ordering to accomodate an
661 /// an edge to be removed from the specified node N from the predecessors
662 /// of the current node M.
663 void RemovePred(SUnit *M, SUnit *N);
665 typedef std::vector<int>::iterator iterator;
666 typedef std::vector<int>::const_iterator const_iterator;
667 iterator begin() { return Index2Node.begin(); }
668 const_iterator begin() const { return Index2Node.begin(); }
669 iterator end() { return Index2Node.end(); }
670 const_iterator end() const { return Index2Node.end(); }
672 typedef std::vector<int>::reverse_iterator reverse_iterator;
673 typedef std::vector<int>::const_reverse_iterator const_reverse_iterator;
674 reverse_iterator rbegin() { return Index2Node.rbegin(); }
675 const_reverse_iterator rbegin() const { return Index2Node.rbegin(); }
676 reverse_iterator rend() { return Index2Node.rend(); }
677 const_reverse_iterator rend() const { return Index2Node.rend(); }